def export_calib_session(session, args, json_name, bcp=False):
if bcp:
vp_json_path = os.path.join(args.bcp_path, 'data', session, '{}.json'.format(json_name))
calib_json_path = os.path.join(args.bcp_path, 'results', session, 'system_{}_{}c.json'.format(json_name, args.conf))
else:
vp_json_path = os.path.join(args.bcs_path, 'dataset', session, '{}.json'.format(json_name))
calib_json_path = os.path.join(args.bcs_path, 'results', session, 'system_{}_{}c.json'.format(json_name, args.conf))
if session == 'S09' or session == 'S10' or session == 'S11':
pp = np.array([640.5, 400.5])
else:
pp = np.array([960.5, 540.5])
print("Starting for session {}".format(session))
with open(vp_json_path, 'r') as f:
vp_data = json.load(f)
if bcp:
vp_data = filter_vp(vp_data)
if 'pred_vars' in vp_data[0].keys():
pred_vars = np.array([item['pred_vars'] for item in vp_data if item['score'] > args.conf])
pred_vars_vp1 = pred_vars[:, :, 0]
pred_vars_vp2 = pred_vars[:, :, 1]
best_scales_vp1 = np.argmin(pred_vars_vp1, axis=-1)
best_scales_vp2 = np.argmin(pred_vars_vp2, axis=-1)
_, counts_vp1 = np.unique(best_scales_vp1, return_counts=True)
_, counts_vp2 = np.unique(best_scales_vp2, return_counts=True)
# print("VP1 best scale counts: ", counts_vp1)
# print("VP2 best scale counts: ", counts_vp2)
else:
counts_vp1 = np.zeros(4)
counts_vp2 = np.zeros(4)
# if 'vp1_var' in vp_data[0].keys():
# scores = np.array([item['score'] / (item['vp1_var'] * item['vp2_var']) for item in vp_data])
# vp1 = np.array([item['vp1'] for item in vp_data])
# vp2 = np.array([item['vp2'] for item in vp_data])
# f = np.sqrt(-np.sum((vp1 - pp[np.newaxis, :]) * (vp2 - pp[np.newaxis, :]), axis=1))
scores = np.array([item['score'] for item in vp_data if item['score'] > args.conf])
vp1 = np.array([item['vp1'] for item in vp_data if item['score'] > args.conf])
vp2 = np.array([item['vp2'] for item in vp_data if item['score'] > args.conf])
f = np.sqrt(-np.sum((vp1 - pp[np.newaxis, :]) * (vp2 - pp[np.newaxis, :]), axis=1))
scores = scores[~np.isnan(f)]
vp1 = vp1[~np.isnan(f)]
vp2 = vp2[~np.isnan(f)]
f = f[~np.isnan(f)]
if args.debug:
show_vps(np.array(vp1), np.array(vp2), session)
med_f = np.nanmedian(f)
m = (vp1[:, 1] - vp2[:, 1]) / (vp1[:, 0] - vp2[:, 0])
b1 = vp1[:, 1] - m * vp1[:, 0]
b2 = vp2[:, 1] - m * vp2[:, 0]
med_m = np.nanmedian(m)
med_k = np.nanmedian(np.concatenate([b1, b2]))
vp1_calib, vp2_calib = get_calib_vp(vp1, med_m, med_k, med_f, pp)
if args.debug:
show_vps(np.array([vp1_calib]), np.array([vp2_calib]), session)
lp1 = np.array([item['lp1'] for item in vp_data if "lp1" in item and item['lp1'] is not None and item['score'] > args.conf])
lp2 = np.array([item['lp2'] for item in vp_data if "lp2" in item and item['lp2'] is not None and item['score'] > args.conf])
if len(lp1) == 0:
calib_dict = get_calib_dict(vp1_calib, vp2_calib, pp)
save(calib_json_path, calib_dict)
else:
dists = []
projector = get_projector(vp1_calib, vp2_calib, pp)
for p1, p2 in zip(lp1, lp2):
ip1 = projector(p1)
ip2 = projector(p2)
dists.append(np.linalg.norm(ip1 - ip2))
med_dist = np.nanmedian(dists)
scale = 0.52 / med_dist
calib_dict = get_calib_dict(vp1_calib, vp2_calib, pp, scale=scale)
save(calib_json_path, calib_dict)
return counts_vp1, counts_vp2
def detect_session(detector, model_dir_name, data_path, session, args):
batch_vp_detector = BatchVPDetectorReg(detector, args)
print("Starting object detection for ", session)
cap = cv2.VideoCapture(
os.path.join(data_path, 'dataset', session, 'video.avi'))
# DO NOT REMOVE OTHERWISE FRAMES WILL NOT SYNC
total_frames = cap.get(cv2.CAP_PROP_FRAME_COUNT)
print("Video loaded!")
if args.mask:
json_path = os.path.join(data_path, 'dataset', session,
'detections_mask.json')
output_json_name = 'VPout_{}_r{}_mask.json'.format(
model_dir_name, args.resume)
else:
json_path = os.path.join(data_path, 'dataset', session,
'detections.json')
output_json_name = 'VPout_{}_r{}.json'.format(model_dir_name,
args.resume)
output_json_path = os.path.join(data_path, 'dataset', session,
output_json_name)
with open(json_path, 'r') as f:
detection_data = json.load(f)
total_box_count = sum([len(item['boxes']) for item in detection_data])
print("Loaded {} bounding boxes for {} frames".format(
total_box_count, len(detection_data)))
box_cnt = 0
start_time = time.time()
for detection_cnt, detection in enumerate(detection_data):
for _ in range(args.skip):
ret, frame = cap.read()
if not ret or frame is None:
break
frame_cnt_orig = cap.get(cv2.CAP_PROP_POS_FRAMES)
frame_cnt = detection['frame_cnt']
if frame_cnt != frame_cnt_orig:
raise Exception(
"Frames from OD do not match frames now! Wrong skip param?")
boxes = detection['boxes']
scores = detection['scores']
if args.mask:
masks = detection['masks']
if args.debug:
show_debug(np.copy(frame), boxes)
for i in range(len(boxes)):
box_cnt += 1
if args.mask:
batch_vp_detector.process(frame,
boxes[i],
scores[i],
frame_cnt=frame_cnt,
mask=masks[i])
else:
batch_vp_detector.process(frame,
boxes[i],
scores[i],
frame_cnt=frame_cnt)
if args.dump_every != 0 and detection_cnt % args.dump_every == 0:
print("Saving at detection ", detection_cnt)
save(output_json_path, batch_vp_detector.output_list)
remaining_seconds = (time.time() - start_time) / (box_cnt + 1) * (
total_box_count - box_cnt)
print('Frame {}, Box: {} / {}, ETA: {}'.format(
frame_cnt, box_cnt, total_box_count,
datetime.timedelta(seconds=remaining_seconds)))
batch_vp_detector.finalize()
print("Saving at box ", box_cnt)
save(output_json_path, batch_vp_detector.output_list)
print("Finished session: {} with {} boxes".format(session,
total_box_count))
def detect_session(detector, model_dir_name, data_path, session, args):
batch_vp_detector = BatchVPDetectorReg(detector, args)
print("Starting vp detection for ", session)
if args.mask:
output_json_name = 'VPout_{}_r{}_mask.json'.format(model_dir_name, args.resume)
json_path = os.path.join(data_path, 'data', session, 'detections_mask.json')
else:
output_json_name = 'VPout_{}_r{}.json'.format(model_dir_name, args.resume)
json_path = os.path.join(data_path, 'data', session, 'detections.json')
with open(json_path, 'r') as f:
detection_data = json.load(f)
detection_data = detection_data[:args.max_frames]
output_json_path = os.path.join(data_path, 'data', session, output_json_name)
total_box_count = sum([len(item['boxes']) for item in detection_data])
print("Loaded {} bounding boxes for {} frames".format(total_box_count, len(detection_data)))
box_cnt = 0
start_time = time.time()
# running_average_frame = np.zeros([1080, 1920, 2], dtype=np.float32)
prev_edge = None
masks = None
for detection_cnt, detection in enumerate(detection_data):
frame_filename = detection['filename']
frame = cv2.imread(os.path.join(data_path, 'frames', session, frame_filename))
frame_cnt = detection['frame_cnt']
boxes = detection['boxes']
scores = detection['scores']
del detection['boxes']
del detection['scores']
box_cnt += len(boxes)
if args.mask:
masks = detection['masks']
del detection['masks']
boxes, scores, masks, prev_edge = filter_boxes_bcp(boxes, scores, frame, prev_edge, masks=masks)
if args.debug:
show_debug(np.copy(frame), boxes)
for i in range(len(boxes)):
if args.mask:
batch_vp_detector.process(frame, boxes[i], scores[i], frame_cnt=frame_cnt, frame_filename=frame_filename, mask=masks[i])
else:
batch_vp_detector.process(frame, boxes[i], scores[i], frame_cnt=frame_cnt, frame_filename=frame_filename)
if args.dump_every != 0 and detection_cnt % args.dump_every == 0:
print("Saving at detection ", detection_cnt)
save(output_json_path, batch_vp_detector.output_list)
remaining_seconds = (time.time() - start_time) / (box_cnt + 1) * (total_box_count - box_cnt)
print('{} : {}, Box: {} / {}, ETA: {}'.format(frame_cnt, frame_filename, box_cnt, total_box_count, datetime.timedelta(seconds=remaining_seconds)))
batch_vp_detector.finalize()
print("Saving at box ", box_cnt)
save(output_json_path, batch_vp_detector.output_list)
print("Finished session: {} with {} boxes".format(session, total_box_count))